A task that once filled Lucas Melo's day now takes three hours: prompt an AI agent, step away, then return to working code. Scaling agents exposes a gap few enterprises measure.
"That's just a couple months of progress," Melo told The Forecast. "It's not magic, but it's also magic."
That productivity leap helps explain why many enterprises are rushing to adopt agentic AI. But scaling AI agents across enterprise applications creates a very different set of challenges from scaling coding agents across an organization. The central problem, Melo said, is measuring agent performance. Familiar benchmarks gauge the reasoning step of generative AI solutions, but they tell organizations little about the resource utilization, latency and success rate of agents that can plan work, call external tools and systems, and carry out actions independently.
Agentic AI performance metrics are the measurements enterprises use to evaluate how AI agents run at scale, capturing task success rate, cost per task, time per task, throughput, agent density and latency rather than reasoning alone.
"Beyond reasoning, the agents are also planning, acting, and then reflecting," Melo said. "That's a multi-step loop. So for enterprises to deploy AI agents at scale, they need to think about agentic AI as an overall system."
Where Are AI Agents Gaining Ground in the Enterprise?
Much of the most visible agentic AI activity remains centered on software development. Developers adopt new technology quickly, Melo said, and coding environments provide a relatively controlled way for organizations to experiment with autonomous systems.
Melo spoke with The Forecast shortly after he and Keegan Sheedy, a cloud solutions architect at Intel, co-authored "Building the Enterprise Environment for Agentic AI," a sponsored article published in MIT Technology Review. In that article, they explain why enterprises will find success with a complete agentic AI environment where agents plan, retrieve, remember and act reliably at scale.
Many enterprises are beginning to look beyond coding assistants, toward AI agents that can automate portions of business processes. Forrester reports that three-quarters of enterprise leaders say their organizations are adopting agentic AI, but only a small minority are running agents in "meaningful production," with true scaled multi-agent systems even rarer. Deloitte says that just 15% of organizations have already scaled "orchestrated, cross-functional multi-agent adoption," although most leaders expect nearly half of their business processes to be redesigned or rebuilt around AI agents within the next few years.
Melo recently spoke with leaders at one organization exploring agentic AI for intelligent document processing. The company handles files in various formats, and leaders wanted to extract information, images and tables to produce standardized executive reports. The plan is to use AI agents to free developers from explicitly programming every step and exception.
"English becomes the programming language," Melo said. "You're not using Python, you're not using HTML. You are programming the agent with English, with a goal and an outcome. And because the agent now has agency and reasoning capabilities, it's much more dynamic and flexible when it comes to doing complex jobs."
Why Do Existing Metrics Fall Short for Agentic AI?
When Melo first spoke with The Forecast in April 2026 at the .NEXT event in Chicago, he saw the early impact of agentic AI rebalancing the ratio of CPU to GPU.
"Agentic AI is driving additional CPU workload due to the need for orchestration and execution of Agentic AI workloads," he said.
The fundamental differences between generative and agentic AI systems can make familiar metrics misleading, Melo said. During Intel testing, engineers looked at CPU utilization to determine how much agentic activity a system could support. But even as agents began to overwhelm IT systems, CPU utilization stayed modest, making that metric a poor indicator of when agentic AI is approaching practical limits.
"We threw eight agents, 10 agents, 20 agents at the system," Melo said. "The CPU utilization was still only at 70%, but the latency started to skyrocket. That's because the execution and acting and planning processes of agentic AI workloads are very bursty."
Similarly, Melo said tools like Terminal-Bench and SWE-Bench focus on the narrow metric of task-success ratio, evaluating whether a model can provide the reasoning ability needed for agents to complete a given task.
"Those tools do nothing to measure how agents are running at scale across the system," he said.
What Should Enterprises Measure to Scale Agentic AI?
"As with any other system, agentic AI brings a new set of workloads," Melo said. "You need to understand what it takes to measure it, and to understand how it's running so that you can tune it and scale it over time."
Melo said Intel has identified six valuable metrics for evaluating agentic performance: task success rate, cost per task, time per task, task throughput, agent density (agents per vCPU) and latency. Together, he said, these metrics go beyond evaluating the model's reasoning capability and instead provide a more comprehensive picture of overall performance and efficiency. They capture how long tasks take, how many agents a fleet can support, how costs change as more agents work simultaneously and what users experience at the end of the execution process. They also give leaders the information needed to decide how to scale systems to support more agents.
"These are metrics that we don't see most enterprises paying attention to yet," Melo said.
"Most are still only focusing on traditional metrics like time to first token, token latency, and tokens per second. Those are all inferencing metrics. Those are necessary, but then we also need to add metrics for the agents themselves."
How Do These Metrics Shape Scalable Agentic AI?
Not every measurement is equally important for every agentic application, Melo said. For a loan approval agent, accuracy may be the leading metric. In customer support, latency might matter most. For the overall system, measurements like cost per task and agent density help organizations better manage total cost of ownership (TCO).
These measurements will continue to matter, Melo said, even if businesses eventually show positive bottom-line impacts from their agentic AI investments, such as tracing $200,000 in new efficiencies to $100,000 in spending.
"If you're not measuring time per task, agent density, and latency, you're missing overall system observability," he said. "You're not going to have enough data points to understand from a TCO perspective if you are maximizing your infrastructure, or if you are out of capacity and you need to scale out."
The answers will grow more important as AI agents spread within enterprises, Melo said.
"AI is evolving daily, and the pace of innovation is just increasing," he said. "If you experienced AI last year, it's orders of magnitude better today, and it's going to probably be orders of magnitude better next month."
"It's constantly evolving, and this is the slowest that it's ever going to be."
Related:
Calvin Hennick is a contributing writer. His work appears in BizTech, Engineering Inc., The Boston Globe Magazine and elsewhere. He is also the author of Once More to the Rodeo: A Memoir. Find him on LinkedIn.
© 2026 Nutanix, Inc. All rights reserved. For additional information and important legal disclaimers, please go here.
Chatter about virtualization's demise has grown louder in recent years, fueled by changes following Broadcom's acquisition of VMware and the rapid rise of containers. The reality inside enterprise data centers tells a different tale, according to Deepak Goel, chief technology officer for cloud-native at Nutanix.
"The narrative that virtualization is no longer necessary is one of the most misleading stories circulating in the market today," Goel explained in his blog post The Race to Fill the Virtualization Gap. "Reality is far more nuanced."
Server virtualization is the software abstraction of physical server resources into virtual machines, allowing multiple workloads to run independently on shared hardware. Despite the rise of containers, it remains foundational to enterprise IT for workload isolation, resource efficiency and operational management.
Gartner called the current moment "the most significant disruption in decades" for server virtualization, driven by changes rolled out following Broadcom's 2023 acquisition of VMware. Many VMware customers questioned whether to continue using the company's hypervisor and software, prompting them to explore other options.
There’s a growing need for managing AI performance and efficiency. It’s forcing IT teams to move beyond juggling separate management systems for their vast VM base and growing containerized applications, pushing them toward a unified, single control plane for both environments, according to Steve Carter, director of product marketing at Nutanix.
“While the industry spent a decade or more moving workloads into VMs, the massive compute requirements of AI has many eager to tap into raw power of bare metal for certain needs.”
This is giving rise to what Carter calls “dual native.”
“It’s the ability to run Kubernetes on virtual machines and on bare metal systems, and manage them as part of a single platform,” Carter said.
He said what sets Nutanix apart is its dual native ability to unify both environments, giving IT teams ultimate flexibility without the usual infrastructure trade-offs. He described a dual native-capable IT platform like a shipping hub that sends packages by ground or air depending on urgency and cost. IT departments can route containerized workloads to VMs or bare-metal servers, where those workloads run without a virtualization layer.
Carter said that having a single platform that can switch between both environments offers several benefits.
"With dual native architecture, you can be intentional about the infrastructure that you use,” he said.
"You get the operational efficiency of virtualization for most standard deployments. Or you can support AI, ML, and other resource-intensive workflows that require constant access to GPU hardware by running containers on bare-metal servers rather than on hypervisors.
“It's the best of both worlds."
AI Infrastructure and Application Expansion
Like they’ve done through other transformative times, Carter said organizations want to maximize their investment while simplifying growing IT complexity and security requirements. Many will prioritize investment in AI capabilities that are flexible and can evolve over time.
McKinsey projects that the IT capacity needed for AI and non-AI workloads could almost triple by 2030, with AI capacity increasing 3.5 times and making up roughly 70% of the total.
By default, most investments in IT resources will flow into cloud native technologies, since about 85% of global IT executives in this year's Enterprise Cloud Index (ECI) survey said they expect app containerization to increase, largely driven by AI.
“Containers in general and Kubernetes in particular are the de facto standard for developing and deploying new applications,” said Dan Ciruli, vice president and general manager for cloud native with Nutanix.
"As a developer, it's so much easier to put your app in a container,” Ciruli said. “It makes testing and deployment really, really simple.”
Cloud service providers such as AWS, Azure, and Google Cloud run Kubernetes on virtual machines by default because it delivers strong isolation and operational consistency at scale, Ciruli explained.
“Using Kubernetes in the cloud is solved,” Ciruli said. “But using Kubernetes on-prem can still be difficult.”
This led Ciruli and his team to build a single control plane that runs Kubernetes anywhere. The Nutanix Kubernetes Platform (NKP) simplifies microservices management in on-prem data centers, in the cloud and on edge devices.
“NKP Full Stack lets enterprises run any combination of virtualized and containerized applications,” Ciruli said.
For most enterprises, running cloud native applications in VMs, containers or bare metal is about choice and flexibility, Carter explained.
“Virtualization is the standard used by cloud providers,” he said. “That’s what powers their Kubernetes offerings. People increasingly want to run some of their specialized containers on bare metal. They want direct GPU access when lower latency matters most.”
The ability to manage across different IT infrastructures is becoming more critical as AI infrastructure and applications scale and evolve alongside existing systems.
"You don't have to compromise," Carter said. "You don't have to choose to virtualize containers or not to. You can do both depending on the scenario."
Editor's note: In this video with HyperFRAME Research, Steve Carter describes Why Nutanix is betting on Bare Metal for AI Workloads. Learn more about Nutanix Enterprise AI and Nutanix Kubernetes Platform solutions.
David Rand is a business and technology reporter whose work has appeared in major publications around the world. He specializes in spotting and digging into what’s coming next–and helping executives in organizations of all sizes know what to do about it.
Ken Kaplan contributed to this story. He is Editor in Chief for The Forecast by Nutanix. Find him on X @kenekaplan and LinkedIn.
© 2026 Nutanix, Inc. All rights reserved. For additional information and important legal disclaimers, please go here.
Related Articles
